Say I start with some <M, ∈> a transitive model of a large fragment of ZFC (say enough to run Łoś' Theorem externally) and a specific set x∈M. Now let's say I'm going to pick some M-ultrafilter U on x. By M-ultrafilter, I mean that U measures the subsets of x which are in M.

My question is, by varying U, how much can I affect the ultrapower of M by U? Let's say I limit myself to a U which are countably complete, so that the ultrapower will be wellfounded. If this question is too vague or broad, I'd welcome any interesting examples of things that are possible or impossible.